evermemory（evermemory 开发）

EverMemory

EverMemory is the deterministic memory plugin for OpenClaw. It gives the agent persistent memory, user understanding, and governed self-improvement without hiding the storage or decision process.

What to do first

• When the user is new to EverMemory, start with onboarding.
• When the user asks to remember something important, store it with an explicit kind.
• When the user asks for prior context, recall before answering.
• When the user asks for debugging, auditing, cleanup, backup, or recovery, use the governance and IO tools instead of guessing.

Core model

Layer 1: Memory

• Store durable facts, decisions, preferences, constraints, lessons, and project context.
• Recall by keyword, structured filters, or hybrid retrieval.
• Archive stale or superseded memories and restore them with review/apply gates.

Layer 2: Understanding

• Build a user profile from explicit statements and stable interaction patterns.
• Track behavior rules and preference hints that can shape future responses.
• Generate session briefings so a new session starts with continuity.

Layer 3: Proactivity

• Extract intent and reflections from interaction history.
• Consolidate duplicate or stale memories.
• Explain why a write, recall, archive, or rule action happened.

Tool map

EverMemory has 16 core capabilities. In the current OpenClaw plugin, 15 are exposed as tool commands, and onboarding is registered as profile_onboard. Smartness exists in the SDK/status layer but is not currently registered as a standalone OpenClaw tool.

Tool usage guidance

evermemory_store

Use for explicit long-term facts. Prefer concise, high-value content and a correct kind.

Example:

{
  "content": "Technical decision: replace Webpack with Vite.",
  "kind": "decision"
}

Store when the user says:

• "记住这个决定"
• "以后按这个偏好来"
• "这个坑以后别再踩"

evermemory_recall

Use before answering when the user asks about prior context, preferences, constraints, or project continuity.

Example:

{
  "query": "Vite migration decision",
  "limit": 5
}

evermemory_status

Use for health checks and operator-style visibility. It returns memory counts, archive counts, profile/rule/reflection state, recent debug activity, and continuity KPIs.

evermemory_briefing

Use at session start or when the user asks for a summary of who they are, current constraints, and active project context.

profile_onboard

Use for first-run setup. Ask the questions, collect answers, then submit them. Do not skip onboarding if no profile exists and the user wants personalized memory behavior.

evermemory_profile

Use to inspect current user understanding. Prefer recompute: true when the user asks for a refreshed profile after many new interactions.

evermemory_rules

Use for behavior rules and guardrails. Prefer read/review paths before mutating rules.

evermemory_explain

Use when the user asks "why did you remember this", "why was this recalled", "why was this archived", or "why did this rule trigger".

evermemory_export and evermemory_import

• Export for backup or migration.
• Import with mode: "review" first.
• Only use apply after the user clearly confirms.

evermemory_review and evermemory_restore

• Review archived memory before restoring.
• Prefer mode: "review" first.
• Restore only the specific IDs the user approves.

evermemory_intent, evermemory_reflect, evermemory_consolidate

Use these as maintenance and self-improvement tools:

• evermemory_intent for intent labeling and routing insight.
• evermemory_reflect for lessons, warnings, and candidate rules.
• evermemory_consolidate for dedupe and stale-memory cleanup.

Recommended workflows

First use

用户: 开始使用 EverMemory
动作: 调用 profile_onboard
结果: 完成初始化问卷，建立基础画像

Remember a decision

用户: 记住我们决定用 Vite 替代 Webpack
动作: 调用 evermemory_store
建议 kind: decision

Recall previous context

用户: 回忆一下我们上次怎么定的
动作: 先调用 evermemory_recall，再基于召回结果回答

Export backup

用户: 导出所有记忆为 JSON
动作: 调用 evermemory_export，并使用 format=json（OpenClaw 注册层）

Recovery

用户: 把之前归档掉的 TypeScript 偏好恢复回来
动作: 先调用 evermemory_review 找候选，再调用 evermemory_restore

Guardrails

• Do not claim a standalone evermemory_smartness tool exists unless the host actually registers it.
• In the current repository, onboarding is profile_onboard, not evermemory_onboard.
• Prefer review before apply for import and restore.
• Recall before answering if the user explicitly asks about previous decisions, preferences, or history.
• Store only durable, high-signal information. Avoid writing transient chatter as memory.
• When a tool returns governed results, explain them plainly instead of exposing raw internals unless the user asks.

Configuration notes

Common environment variables for semantic retrieval:

• EVERMEMORY_EMBEDDING_PROVIDER: local, openai, or none
• EVERMEMORY_LOCAL_MODEL: local embedding model, default Xenova/all-MiniLM-L6-v2
• OPENAI_API_KEY: required when the embedding provider uses OpenAI

Common plugin config fields:

• databasePath
• bootTokenBudget
• maxRecall
• debugEnabled
• semantic.enabled
• semantic.maxCandidates
• semantic.minScore
• intent.useLLM
• intent.fallbackHeuristics